Collective Emergent Alignment: Measuring and Amplifying Interaction-Induced Reasoning Alignment in Multi-Agent LLMs

The researchers propose the REMAP (Reference-Entailment-Matrix for Agentic Process) framework to quantify interaction-induced reasoning alignment in multi-agent LLM systems, and find that across multiple model families and two multi-hop reasoning benchmarks, the alignment gain from interaction is generally positive, and is especially pronounced on reasoning content that requires integrating evidence distributed across different agents. The framework measures this effect via Semantic Coverage and an interaction-specific alignment gain; the gain correlates consistently with independently measured peer information recovery, and further analysis finds heterogeneity effects that vary with composition and asymmetric information transfer across agent capability levels. Based on these findings, the authors propose a fact-level multi-agent communication framework that replaces full reasoning trajectory exchange with compact, gap-aware, selectively routed information; experiments show that compact fact exchange improves reasoning alignment and task performance while reducing communication overhead, whereas more selective communication produces trade-offs among acquiring distributed information, preserving local content, and communication efficiency.